Retrieval on human-curated retrieval benchmark
60Top-1 RecallMiniLM-L6
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| MiniLM-L6Param. Count=22.7M, strategy=Cat Only, details=strict categorical matching only2025.12 | 60 | 72.73 | 76.36 | 80 | |
| MiniLM-L6Param. Count=22.7M, strategy=Cat + Sem Rerank, details=categorical matching followed by semantic re-ranking2025.12 | 58.18 | 74.55 | 78.18 | 80 | |
| OpenAI ada-002Param. Count=N/A2025.12 | 52.7 | 63.6 | 67.3 | 70.9 | |
| MiniLM-L6Param. Count=22.7M, strategy=Default, details=semantic filtering followed by categorical checks2025.12 | 50.91 | 60 | 67.27 | 70.91 | |
| OpenAI 3 LargeParam. Count=N/A2025.12 | 49.1 | 65.5 | 67.3 | 72.7 | |
| Qwen2 0.6BParam. Count=600M2025.12 | 38.2 | 58.2 | 61.8 | 65.5 | |
| Qwen2 4BParam. Count=4000M2025.12 | 30.9 | 50.9 | 50.9 | 63.6 | |
| Qwen2 1.5BParam. Count=1500M2025.12 | 30.9 | — | 54.5 | 60 | |
| SciBERTParam. Count=110M2025.12 | 23.6 | 29.1 | 38.2 | 58.2 | |
| RoBERTa LargeParam. Count=355M2025.12 | 20 | 36.4 | 45.5 | 50.9 | |
| MiniLM-L6Param. Count=22.7M, strategy=Sem, details=semantic matching only2025.12 | 3.64 | 10.91 | 12.73 | 14.55 |